Data Resampling and Ensemble Learning for Vehicle Class and Orientation Detection

Vehicle class and orientation detection is fundamental to different tasks of autonomous driving such as traffic flow estimation, vehicle re-identification, etc. Recent researches using Convolutional Neural Networks (CNNs) or Vision-based Transformers for object detection tasks require a large number of images with annotations. The collection of a large amount of high-quality dataset is expensive and time-consuming. Synthetic images obtained from driving simulators allow the generation of a large synthetic dataset under varying conditions. This paper proposes YOLOv7-based neural networks for vehicle class and orientation detection that train on the synthetic dataset made available as part of the Vehicle class and Orientation Detection Challenge 2022 of IEEE BigData 2022. Our proposed approaches included the use of data resampling methods to address data imbalance in the dataset and ensemble learning to optimize the model performance. Trained models were thoroughly evaluated on real-world test datasets. Our proposed approach achieved weighted mean average precision of 0.497 allowing us to win the challenge.